Images with cut-out regions (rects, circles, corners, strips, irregular tears) and matching masks for testing inpainting, generative fill, content-aware fill, and object-removal pipelines.
Black-and-white damage mask for the crease restoration case (white = damaged). Use with the damaged input and clean reference.
Black-and-white damage mask for the scratch restoration case (white = damaged). Use with the damaged input and clean reference.
Black-and-white damage mask for the dust restoration case (white = damaged). Use with the damaged input and clean reference.
Black-and-white damage mask for the water-stain restoration case (white = damaged). Use with the damaged input and clean reference.
Black-and-white damage mask for the torn-corner restoration case (white = damaged). Use with the damaged input and clean reference.
Black-and-white damage mask for the horizontal-cut restoration case (white = damaged). Use with the damaged input and clean reference.
Clean fruit still-life reference for the inpaint cut suite. Each cut variant and mask in this group points back to this file.
Missing-region (centre-rect) input for inpainting / content-aware fill tests. Use with the matching mask and clean reference.
White = missing region for the centre-rect cut. Feed this mask with the cut input to an inpainting tool.
Missing-region (centre-circle) input for inpainting / content-aware fill tests. Use with the matching mask and clean reference.
White = missing region for the centre-circle cut. Feed this mask with the cut input to an inpainting tool.
Missing-region (corner) input for inpainting / content-aware fill tests. Use with the matching mask and clean reference.
White = missing region for the corner cut. Feed this mask with the cut input to an inpainting tool.
Missing-region (vertical-strip) input for inpainting / content-aware fill tests. Use with the matching mask and clean reference.
White = missing region for the vertical-strip cut. Feed this mask with the cut input to an inpainting tool.
Missing-region (irregular-tear) input for inpainting / content-aware fill tests. Use with the matching mask and clean reference.
White = missing region for the irregular-tear cut. Feed this mask with the cut input to an inpainting tool.
Missing-region (edge-crop) input for inpainting / content-aware fill tests. Use with the matching mask and clean reference.
White = missing region for the edge-crop cut. Feed this mask with the cut input to an inpainting tool.
Missing-region (transparent-hole) input for inpainting / content-aware fill tests. Use with the matching mask and clean reference.
White = missing region for the transparent-hole cut. Feed this mask with the cut input to an inpainting tool.
A clean 512x512 crop of head and shoulders portrait of a black woman with locs, plain terracotta studio background, taken from the published plate a-av-f3.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of head and shoulders portrait of a black woman with locs, plain terracotta studio background, with a foreign object composited over 6.15% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over head and shoulders portrait of a black woman with locs, plain terracotta studio background, as an 8-bit mask covering 6.15% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Head and shoulders portrait of a black woman with locs, plain terracotta studio background, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 38.409/255 and from the ground-truth plate by 5.719/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 3.157/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of head and shoulders portrait of a smiling black man against a plain grey studio background, taken from the published plate a-avatar-m1.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of head and shoulders portrait of a smiling black man against a plain grey studio background, with a foreign object composited over 5.29% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over head and shoulders portrait of a smiling black man against a plain grey studio background, as an 8-bit mask covering 5.29% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Head and shoulders portrait of a smiling black man against a plain grey studio background, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 54.279/255 and from the ground-truth plate by 10.905/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 4.019/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of moody album cover artwork, lone figure silhouette against fog, muted palette, taken from the published plate c-album-moody.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of moody album cover artwork, lone figure silhouette against fog, muted palette, with a foreign object composited over 10.96% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over moody album cover artwork, lone figure silhouette against fog, muted palette, as an 8-bit mask covering 10.96% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Moody album cover artwork, lone figure silhouette against fog, muted palette, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 46.628/255 and from the ground-truth plate by 1.685/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 2.574/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of an open paper accounts ledger with ruled columns, fountain pen resting on the page, window light, taken from the published plate d-ledger.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of an open paper accounts ledger with ruled columns, fountain pen resting on the page, window light, with a foreign object composited over 5.60% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over an open paper accounts ledger with ruled columns, fountain pen resting on the page, window light, as an 8-bit mask covering 5.60% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
An open paper accounts ledger with ruled columns, fountain pen resting on the page, window light, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 106.971/255 and from the ground-truth plate by 21.952/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 3.353/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of dark dungeon interior video game environment, torchlight, stone arches, taken from the published plate g-dungeon.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of dark dungeon interior video game environment, torchlight, stone arches, with a foreign object composited over 15.76% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over dark dungeon interior video game environment, torchlight, stone arches, as an 8-bit mask covering 15.76% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Dark dungeon interior video game environment, torchlight, stone arches, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 62.511/255 and from the ground-truth plate by 24.609/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 7.771/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of low poly stylised island with trees and water, taken from the published plate g-lowpoly-isle.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of low poly stylised island with trees and water, with a foreign object composited over 6.50% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over low poly stylised island with trees and water, as an 8-bit mask covering 6.50% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Low poly stylised island with trees and water, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 55.589/255 and from the ground-truth plate by 75.537/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 5.011/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of arcade racing game track through a canyon, motion blur, taken from the published plate g-racing-track.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of arcade racing game track through a canyon, motion blur, with a foreign object composited over 15.07% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over arcade racing game track through a canyon, motion blur, as an 8-bit mask covering 15.07% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Arcade racing game track through a canyon, motion blur, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 32.313/255 and from the ground-truth plate by 12.923/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 4.956/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of steel truss bridge over a river, clear daylight, taken from the published plate l-bridge.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of steel truss bridge over a river, clear daylight, with a foreign object composited over 7.24% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over steel truss bridge over a river, clear daylight, as an 8-bit mask covering 7.24% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Steel truss bridge over a river, clear daylight, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 65.894/255 and from the ground-truth plate by 41.698/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 8.333/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of concrete spiral stairwell seen from below, strong geometry, daylight, taken from the published plate l-stairwell.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of concrete spiral stairwell seen from below, strong geometry, daylight, with a foreign object composited over 11.73% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over concrete spiral stairwell seen from below, strong geometry, daylight, as an 8-bit mask covering 11.73% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Concrete spiral stairwell seen from below, strong geometry, daylight, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 65.853/255 and from the ground-truth plate by 31.839/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 5.063/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of a smiling woman with short grey hair in a grey apron, arms folded, standing at a bistro counter below chalkboard menus with bread baskets behind her, taken from the published plate owner-restaurant.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of a smiling woman with short grey hair in a grey apron, arms folded, standing at a bistro counter below chalkboard menus with bread baskets behind her, with a foreign object composited over 10.86% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over a smiling woman with short grey hair in a grey apron, arms folded, standing at a bistro counter below chalkboard menus with bread baskets behind her, as an 8-bit mask covering 10.86% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
A smiling woman with short grey hair in a grey apron, arms folded, standing at a bistro counter below chalkboard menus with bread baskets behind her, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 118.767/255 and from the ground-truth plate by 12.881/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 14.477/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of an electrician in work clothes checking a consumer unit in a domestic hallway, tool bag open on the floor, natural light, taken from the published plate owner-trades.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of an electrician in work clothes checking a consumer unit in a domestic hallway, tool bag open on the floor, natural light, with a foreign object composited over 4.99% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over an electrician in work clothes checking a consumer unit in a domestic hallway, tool bag open on the floor, natural light, as an 8-bit mask covering 4.99% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
An electrician in work clothes checking a consumer unit in a domestic hallway, tool bag open on the floor, natural light, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 72.598/255 and from the ground-truth plate by 18.235/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 4.619/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of a wheelchair user working at an accessible office desk with two monitors and printed papers, open plan office, soft daylight, taken from the published plate p-desk-chair.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of a wheelchair user working at an accessible office desk with two monitors and printed papers, open plan office, soft daylight, with a foreign object composited over 9.28% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over a wheelchair user working at an accessible office desk with two monitors and printed papers, open plan office, soft daylight, as an 8-bit mask covering 9.28% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
A wheelchair user working at an accessible office desk with two monitors and printed papers, open plan office, soft daylight, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 65.106/255 and from the ground-truth plate by 59.559/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 5.553/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of senior south asian woman florist wrapping stems in paper, bucket of blooms, shopfront light, taken from the published plate p-florist-sen.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of senior south asian woman florist wrapping stems in paper, bucket of blooms, shopfront light, with a foreign object composited over 5.02% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over senior south asian woman florist wrapping stems in paper, bucket of blooms, shopfront light, as an 8-bit mask covering 5.02% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Senior south asian woman florist wrapping stems in paper, bucket of blooms, shopfront light, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 48.638/255 and from the ground-truth plate by 39.582/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 5.552/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of black woman mechanic in overalls under a raised car, work lamp, garage interior, taken from the published plate p-mechanic-w.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of black woman mechanic in overalls under a raised car, work lamp, garage interior, with a foreign object composited over 10.24% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over black woman mechanic in overalls under a raised car, work lamp, garage interior, as an 8-bit mask covering 10.24% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Black woman mechanic in overalls under a raised car, work lamp, garage interior, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 49.413/255 and from the ground-truth plate by 57.367/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 7.292/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of a plus size white woman in her fifties behind the counter of a small hardware shop, shelves of tools behind her, shopfront daylight, taken from the published plate p-shop-plus.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of a plus size white woman in her fifties behind the counter of a small hardware shop, shelves of tools behind her, shopfront daylight, with a foreign object composited over 7.87% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over a plus size white woman in her fifties behind the counter of a small hardware shop, shelves of tools behind her, shopfront daylight, as an 8-bit mask covering 7.87% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
A plus size white woman in her fifties behind the counter of a small hardware shop, shelves of tools behind her, shopfront daylight, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 47.083/255 and from the ground-truth plate by 59.823/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 4.116/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of latina teacher beside a whiteboard talking to seated adult students, classroom, window light, taken from the published plate p-teacher-cls.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of latina teacher beside a whiteboard talking to seated adult students, classroom, window light, with a foreign object composited over 16.78% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over latina teacher beside a whiteboard talking to seated adult students, classroom, window light, as an 8-bit mask covering 16.78% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Latina teacher beside a whiteboard talking to seated adult students, classroom, window light, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 43.944/255 and from the ground-truth plate by 42.892/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 4.889/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of open laptop on a plain white studio background, three quarter, soft light, blank screen, taken from the published plate pr-laptop.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of open laptop on a plain white studio background, three quarter, soft light, blank screen, with a foreign object composited over 8.33% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over open laptop on a plain white studio background, three quarter, soft light, blank screen, as an 8-bit mask covering 8.33% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Open laptop on a plain white studio background, three quarter, soft light, blank screen, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 53.18/255 and from the ground-truth plate by 4.38/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 10.44/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of wristwatch on a plain grey studio background, soft light, taken from the published plate pr-watch.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of wristwatch on a plain grey studio background, soft light, with a foreign object composited over 9.31% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over wristwatch on a plain grey studio background, soft light, as an 8-bit mask covering 9.31% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Wristwatch on a plain grey studio background, soft light, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 35.125/255 and from the ground-truth plate by 7.094/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 8.446/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of an empty modern open plan office with desks and monitors, late afternoon sun through blinds, no people, taken from the published plate s-office.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of an empty modern open plan office with desks and monitors, late afternoon sun through blinds, no people, with a foreign object composited over 4.75% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over an empty modern open plan office with desks and monitors, late afternoon sun through blinds, no people, as an 8-bit mask covering 4.75% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
An empty modern open plan office with desks and monitors, late afternoon sun through blinds, no people, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 20.037/255 and from the ground-truth plate by 16.793/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 4.342/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of group of road cyclists on a country road, daylight, taken from the published plate sp-cyclists.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of group of road cyclists on a country road, daylight, with a foreign object composited over 9.72% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over group of road cyclists on a country road, daylight, as an 8-bit mask covering 9.72% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Group of road cyclists on a country road, daylight, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 44.39/255 and from the ground-truth plate by 13.229/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 6.53/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of swimmer mid stroke in an indoor pool lane, pool lighting, taken from the published plate sp-swimmer.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of swimmer mid stroke in an indoor pool lane, pool lighting, with a foreign object composited over 10.91% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over swimmer mid stroke in an indoor pool lane, pool lighting, as an 8-bit mask covering 10.91% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Swimmer mid stroke in an indoor pool lane, pool lighting, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 53.16/255 and from the ground-truth plate by 54.228/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 4.358/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of smooth abstract gradient, deep teal to midnight blue, subtle noise, taken from the published plate t-grad-cool.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of smooth abstract gradient, deep teal to midnight blue, subtle noise, with a foreign object composited over 16.26% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over smooth abstract gradient, deep teal to midnight blue, subtle noise, as an 8-bit mask covering 16.26% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Smooth abstract gradient, deep teal to midnight blue, subtle noise, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 80.562/255 and from the ground-truth plate by 6.576/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 2.895/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of smooth abstract gradient, warm amber to deep rose, soft film grain, taken from the published plate t-grad-warm.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of smooth abstract gradient, warm amber to deep rose, soft film grain, with a foreign object composited over 12.70% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over smooth abstract gradient, warm amber to deep rose, soft film grain, as an 8-bit mask covering 12.70% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Smooth abstract gradient, warm amber to deep rose, soft film grain, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 59.829/255 and from the ground-truth plate by 7.845/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 9.08/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of soft blurred abstract background, muted sage green, very low detail, taken from the published plate t-ph-soft3.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of soft blurred abstract background, muted sage green, very low detail, with a foreign object composited over 3.85% of the frame as two disconnected regions - the case a reader that keeps only the largest connected component, or takes the bounding box of both, gets wrong. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over soft blurred abstract background, muted sage green, very low detail, as an 8-bit mask covering 3.85% of the frame as two disconnected regions - the case a reader that keeps only the largest connected component, or takes the bounding box of both, gets wrong. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Soft blurred abstract background, muted sage green, very low detail, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 41.655/255 and from the ground-truth plate by 3.338/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 2.982/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of close up indigo denim fabric weave, taken from the published plate t-tex-denim.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of close up indigo denim fabric weave, with a foreign object composited over 7.10% of the frame as two disconnected regions - the case a reader that keeps only the largest connected component, or takes the bounding box of both, gets wrong. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over close up indigo denim fabric weave, as an 8-bit mask covering 7.10% of the frame as two disconnected regions - the case a reader that keeps only the largest connected component, or takes the bounding box of both, gets wrong. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Close up indigo denim fabric weave, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 61.565/255 and from the ground-truth plate by 12.504/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 10.632/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of city bicycle leaning against a brick wall, soft daylight, taken from the published plate v-bike-city.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of city bicycle leaning against a brick wall, soft daylight, with a foreign object composited over 7.55% of the frame as two disconnected regions - the case a reader that keeps only the largest connected component, or takes the bounding box of both, gets wrong. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over city bicycle leaning against a brick wall, soft daylight, as an 8-bit mask covering 7.55% of the frame as two disconnected regions - the case a reader that keeps only the largest connected component, or takes the bounding box of both, gets wrong. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
City bicycle leaning against a brick wall, soft daylight, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 53.072/255 and from the ground-truth plate by 26.044/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 7.882/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of white electric hatchback at a charging point, urban car park, taken from the published plate v-car-ev.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of white electric hatchback at a charging point, urban car park, with a foreign object composited over 4.31% of the frame as two disconnected regions - the case a reader that keeps only the largest connected component, or takes the bounding box of both, gets wrong. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over white electric hatchback at a charging point, urban car park, as an 8-bit mask covering 4.31% of the frame as two disconnected regions - the case a reader that keeps only the largest connected component, or takes the bounding box of both, gets wrong. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
White electric hatchback at a charging point, urban car park, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 89.979/255 and from the ground-truth plate by 38.547/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 4.784/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of yellow forklift in a warehouse aisle, industrial lighting, taken from the published plate v-forklift.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of yellow forklift in a warehouse aisle, industrial lighting, with a foreign object composited over 6.26% of the frame as two disconnected regions - the case a reader that keeps only the largest connected component, or takes the bounding box of both, gets wrong. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over yellow forklift in a warehouse aisle, industrial lighting, as an 8-bit mask covering 6.26% of the frame as two disconnected regions - the case a reader that keeps only the largest connected component, or takes the bounding box of both, gets wrong. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Yellow forklift in a warehouse aisle, industrial lighting, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 54.408/255 and from the ground-truth plate by 33.846/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 7.135/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of passenger aircraft at an airport gate, overcast sky, taken from the published plate v-plane-gate.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of passenger aircraft at an airport gate, overcast sky, with a foreign object composited over 14.90% of the frame as two disconnected regions - the case a reader that keeps only the largest connected component, or takes the bounding box of both, gets wrong. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over passenger aircraft at an airport gate, overcast sky, as an 8-bit mask covering 14.90% of the frame as two disconnected regions - the case a reader that keeps only the largest connected component, or takes the bounding box of both, gets wrong. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Passenger aircraft at an airport gate, overcast sky, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 84.008/255 and from the ground-truth plate by 64.23/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 7.224/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of articulated lorry at a distribution depot loading bay, morning light, taken from the published plate v-truck-artic.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of articulated lorry at a distribution depot loading bay, morning light, with a foreign object composited over 8.83% of the frame as one long thin region, where growing the mask by 12 pixels more than doubles the area it covers. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over articulated lorry at a distribution depot loading bay, morning light, as an 8-bit mask covering 8.83% of the frame as one long thin region, where growing the mask by 12 pixels more than doubles the area it covers. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Articulated lorry at a distribution depot loading bay, morning light, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 55.181/255 and from the ground-truth plate by 63.287/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 6.781/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of head and shoulders portrait of a young androgynous person with short pink hair, plain lilac background, taken from the published plate a-av-nb1.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of head and shoulders portrait of a young androgynous person with short pink hair, plain lilac background, with a foreign object composited over 5.08% of the frame as one long thin region, where growing the mask by 12 pixels more than doubles the area it covers. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over head and shoulders portrait of a young androgynous person with short pink hair, plain lilac background, as an 8-bit mask covering 5.08% of the frame as one long thin region, where growing the mask by 12 pixels more than doubles the area it covers. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Head and shoulders portrait of a young androgynous person with short pink hair, plain lilac background, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 51.809/255 and from the ground-truth plate by 7.091/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 2.995/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of stylised fantasy knight character render, full body, neutral grey backdrop, taken from the published plate g-char-knight.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of stylised fantasy knight character render, full body, neutral grey backdrop, with a foreign object composited over 10.66% of the frame as one long thin region, where growing the mask by 12 pixels more than doubles the area it covers. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over stylised fantasy knight character render, full body, neutral grey backdrop, as an 8-bit mask covering 10.66% of the frame as one long thin region, where growing the mask by 12 pixels more than doubles the area it covers. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Stylised fantasy knight character render, full body, neutral grey backdrop, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 52.182/255 and from the ground-truth plate by 20.634/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 6.795/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of science fiction video game corridor, emissive panels, volumetric light, taken from the published plate g-scifi-corr.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of science fiction video game corridor, emissive panels, volumetric light, with a foreign object composited over 10.89% of the frame as a region touching the frame, so the inpainter has no surrounding context on one side. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over science fiction video game corridor, emissive panels, volumetric light, as an 8-bit mask covering 10.89% of the frame as a region touching the frame, so the inpainter has no surrounding context on one side. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Science fiction video game corridor, emissive panels, volumetric light, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 53.065/255 and from the ground-truth plate by 38.959/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 3.258/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of narrow european old town street, cobbles, receding , daylight, taken from the published plate l-street-eu.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of narrow european old town street, cobbles, receding , daylight, with a foreign object composited over 9.47% of the frame as a region touching the frame, so the inpainter has no surrounding context on one side. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over narrow european old town street, cobbles, receding , daylight, as an 8-bit mask covering 9.47% of the frame as a region touching the frame, so the inpainter has no surrounding context on one side. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Narrow european old town street, cobbles, receding , daylight, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 62.106/255 and from the ground-truth plate by 38.069/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 11.016/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of a person's hands signing a document at a desk, pen held over spread paperwork, a wristwatch visible at the cuff, taken from the published plate office-desk.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of a person's hands signing a document at a desk, pen held over spread paperwork, a wristwatch visible at the cuff, with a foreign object composited over 4.54% of the frame as a region touching the frame, so the inpainter has no surrounding context on one side. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over a person's hands signing a document at a desk, pen held over spread paperwork, a wristwatch visible at the cuff, as an 8-bit mask covering 4.54% of the frame as a region touching the frame, so the inpainter has no surrounding context on one side. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
A person's hands signing a document at a desk, pen held over spread paperwork, a wristwatch visible at the cuff, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 158.044/255 and from the ground-truth plate by 7.06/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 3.463/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of middle aged white male butcher in striped apron behind a counter, hanging scales, tiled wall, taken from the published plate p-butcher-mid.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of middle aged white male butcher in striped apron behind a counter, hanging scales, tiled wall, with a foreign object composited over 4.61% of the frame as a region touching the frame, so the inpainter has no surrounding context on one side. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over middle aged white male butcher in striped apron behind a counter, hanging scales, tiled wall, as an 8-bit mask covering 4.61% of the frame as a region touching the frame, so the inpainter has no surrounding context on one side. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Middle aged white male butcher in striped apron behind a counter, hanging scales, tiled wall, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 43.785/255 and from the ground-truth plate by 23.557/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 6.11/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of five colleagues of different ages and ethnicities around a meeting table with printed reports and a laptop, office daylight, taken from the published plate p-meeting-mix.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of five colleagues of different ages and ethnicities around a meeting table with printed reports and a laptop, office daylight, with a foreign object composited over 16.00% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over five colleagues of different ages and ethnicities around a meeting table with printed reports and a laptop, office daylight, as an 8-bit mask covering 16.00% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Five colleagues of different ages and ethnicities around a meeting table with printed reports and a laptop, office daylight, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 82.621/255 and from the ground-truth plate by 56.593/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 6.934/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of canvas backpack on a plain grey studio background, three quarter, soft light, taken from the published plate pr-backpack.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of canvas backpack on a plain grey studio background, three quarter, soft light, with a foreign object composited over 15.42% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over canvas backpack on a plain grey studio background, three quarter, soft light, as an 8-bit mask covering 15.42% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Canvas backpack on a plain grey studio background, three quarter, soft light, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 53.202/255 and from the ground-truth plate by 43.265/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 9.491/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of empty stainless commercial kitchen, clean surfaces, overhead lighting, wide angle, taken from the published plate s-kitchen-clean.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of empty stainless commercial kitchen, clean surfaces, overhead lighting, wide angle, with a foreign object composited over 11.08% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over empty stainless commercial kitchen, clean surfaces, overhead lighting, wide angle, as an 8-bit mask covering 11.08% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Empty stainless commercial kitchen, clean surfaces, overhead lighting, wide angle, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 86.716/255 and from the ground-truth plate by 51.166/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 5.917/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of amateur football match on a grass pitch, players contesting the ball, daylight, taken from the published plate sp-football.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of amateur football match on a grass pitch, players contesting the ball, daylight, with a foreign object composited over 16.40% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over amateur football match on a grass pitch, players contesting the ball, daylight, as an 8-bit mask covering 16.40% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Amateur football match on a grass pitch, players contesting the ball, daylight, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 50.428/255 and from the ground-truth plate by 59.266/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 4.641/255, which is the full-frame VAE round trip and not an edit.
A clean 512x512 crop of small fishing boat moored in a harbour, morning light, taken from the published plate v-boat-harbour.png before anything was added to it. This is the ANSWER KEY for its group: the object in the source file was composited onto this image, so this is exactly what was behind it. Nothing else in the group came from a second tool's guess.
The same view of small fishing boat moored in a harbour, morning light, with a foreign object composited over 5.61% of the frame as a single organic region, neither a rectangle nor an ellipse. This is the file an object-removal tool is given. Outside the mask it is byte-identical to the clean plate beside it, so any difference a tool leaves there is damage it did rather than content it was handed.
The exact footprint of the object sitting over small fishing boat moored in a harbour, morning light, as an 8-bit mask covering 5.61% of the frame as a single organic region, neither a rectangle nor an ellipse. It holds only the values 0 and 255. The footprint is what DREW the object, so it is ground truth by construction rather than a segmentation of it. Hard-edged on purpose: a feathered edge has no exact footprint, and the exactness is the point of shipping it.
Small fishing boat moored in a harbour, morning light, with the object taken back out by Stable Diffusion 1.5 inpainting and the gap reconstructed from the surrounding context alone - the masked latents are erased before sampling, so the model never saw what it was painting over. Inside the mask it differs from the source by 30.711/255 and from the ground-truth plate by 25.736/255; the second number is NOT expected to be small, because an inpainter invents plausible content rather than recovering what was there. Beyond a 16-pixel ring around the mask the frame changes by only 7.026/255, which is the full-frame VAE round trip and not an edit.
A 512x512 view of head and shoulders portrait of an older white woman with short grey hair, plain cream background - the frame an outpainting tool is given, cropped from the published plate nss-a-av-f4_00001_.png. Its companion expands it to 768x512 with 128 pixels added to the left and right, the symmetric case, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of head and shoulders portrait of an older white woman with short grey hair, plain cream background, expanded to 768x512 - 128 pixels added to the left and right, the symmetric case - so 33.3% of this frame is invented canvas. The original sits at pixels 128,0 to 640,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 5.231/255, the VAE round trip, but over the whole rectangle by 6.293/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of head and shoulders portrait of an east asian woman with glasses, plain warm grey background - the frame an outpainting tool is given, cropped from the published plate nss-a-av-f5_00001_.png. Its companion expands it to 512x768 with 128 pixels added above and below, the symmetric case turned on its side, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of head and shoulders portrait of an east asian woman with glasses, plain warm grey background, expanded to 512x768 - 128 pixels added above and below, the symmetric case turned on its side - so 33.3% of this frame is invented canvas. The original sits at pixels 0,128 to 512,640. Compare it INSET by 40 pixels: over that core it differs from the original by only 5.482/255, the VAE round trip, but over the whole rectangle by 6.4/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of head and shoulders portrait of a bearded white man in his thirties, plain olive studio background - the frame an outpainting tool is given, cropped from the published plate nss-a-av-m3_00001_.png. Its companion expands it to 704x512 with 192 pixels added to the right alone, so the original is no longer centred, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of head and shoulders portrait of a bearded white man in his thirties, plain olive studio background, expanded to 704x512 - 192 pixels added to the right alone, so the original is no longer centred - so 27.3% of this frame is invented canvas. The original sits at pixels 0,0 to 512,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 5.277/255, the VAE round trip, but over the whole rectangle by 6.19/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of head and shoulders portrait of a south asian man in his fifties, plain slate studio background - the frame an outpainting tool is given, cropped from the published plate nss-a-av-m4_00001_.png. Its companion expands it to 704x512 with 192 pixels added to the left alone, so the original's origin moves, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of head and shoulders portrait of a south asian man in his fifties, plain slate studio background, expanded to 704x512 - 192 pixels added to the left alone, so the original's origin moves - so 27.3% of this frame is invented canvas. The original sits at pixels 192,0 to 704,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 6.007/255, the VAE round trip, but over the whole rectangle by 5.68/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of head and shoulders portrait of a smiling white woman against a plain blue studio background - the frame an outpainting tool is given, cropped from the published plate nss-a-avatar-f1_00001_.png. Its companion expands it to 512x704 with 192 pixels added below alone, the one case where the origin does NOT move, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of head and shoulders portrait of a smiling white woman against a plain blue studio background, expanded to 512x704 - 192 pixels added below alone, the one case where the origin does NOT move - so 27.3% of this frame is invented canvas. The original sits at pixels 0,0 to 512,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 4.251/255, the VAE round trip, but over the whole rectangle by 4.956/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of head and shoulders portrait of a young hijabi woman against a plain neutral studio background - the frame an outpainting tool is given, cropped from the published plate nss-a-avatar-f2_00001_.png. Its companion expands it to 704x704 with 96 pixels added on all four sides, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of head and shoulders portrait of a young hijabi woman against a plain neutral studio background, expanded to 704x704 - 96 pixels added on all four sides - so 47.1% of this frame is invented canvas. The original sits at pixels 96,96 to 608,608. Compare it INSET by 40 pixels: over that core it differs from the original by only 5.267/255, the VAE round trip, but over the whole rectangle by 9.706/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of head and shoulders portrait of an older east asian man against a plain warm studio background - the frame an outpainting tool is given, cropped from the published plate nss-a-avatar-m2_00001_.png. Its companion expands it to 768x640 with a different amount on every side, which is the case a tool assuming symmetry fails, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of head and shoulders portrait of an older east asian man against a plain warm studio background, expanded to 768x640 - a different amount on every side, which is the case a tool assuming symmetry fails - so 46.7% of this frame is invented canvas. The original sits at pixels 64,32 to 576,544. Compare it INSET by 40 pixels: over that core it differs from the original by only 5.021/255, the VAE round trip, but over the whole rectangle by 8.184/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of abstract album cover artwork, bold overlapping colour shapes, grainy print texture - the frame an outpainting tool is given, cropped from the published plate nss-c-album-abs_00001_.png. Its companion expands it to 768x512 with 128 pixels added to the left and right, the symmetric case, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of abstract album cover artwork, bold overlapping colour shapes, grainy print texture, expanded to 768x512 - 128 pixels added to the left and right, the symmetric case - so 33.3% of this frame is invented canvas. The original sits at pixels 128,0 to 640,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 6.331/255, the VAE round trip, but over the whole rectangle by 10.141/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of retro album cover artwork, sunset gradient and geometric lines, 1980s print feel - the frame an outpainting tool is given, cropped from the published plate nss-c-album-retro_00001_.png. Its companion expands it to 512x768 with 128 pixels added above and below, the symmetric case turned on its side, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of retro album cover artwork, sunset gradient and geometric lines, 1980s print feel, expanded to 512x768 - 128 pixels added above and below, the symmetric case turned on its side - so 33.3% of this frame is invented canvas. The original sits at pixels 0,128 to 512,640. Compare it INSET by 40 pixels: over that core it differs from the original by only 9.476/255, the VAE round trip, but over the whole rectangle by 10.985/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of book cover artwork, weathered antique map and brass instruments on wood - the frame an outpainting tool is given, cropped from the published plate nss-c-book-hist_00001_.png. Its companion expands it to 704x512 with 192 pixels added to the right alone, so the original is no longer centred, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of book cover artwork, weathered antique map and brass instruments on wood, expanded to 704x512 - 192 pixels added to the right alone, so the original is no longer centred - so 27.3% of this frame is invented canvas. The original sits at pixels 0,0 to 512,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 7.793/255, the VAE round trip, but over the whole rectangle by 8.625/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of book cover artwork, misty mountain ridge at dawn, layered depth, serene - the frame an outpainting tool is given, cropped from the published plate nss-c-book-nature_00001_.png. Its companion expands it to 704x512 with 192 pixels added to the left alone, so the original's origin moves, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of book cover artwork, misty mountain ridge at dawn, layered depth, serene, expanded to 704x512 - 192 pixels added to the left alone, so the original's origin moves - so 27.3% of this frame is invented canvas. The original sits at pixels 192,0 to 704,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 2.563/255, the VAE round trip, but over the whole rectangle by 3.277/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of a delivery note on a clipboard held against a van door, outdoor overcast light, slight angle - the frame an outpainting tool is given, cropped from the published plate nss-d-clipboard_00001_.png. Its companion expands it to 512x704 with 192 pixels added above alone, so every row index shifts, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of a delivery note on a clipboard held against a van door, outdoor overcast light, slight angle, expanded to 512x704 - 192 pixels added above alone, so every row index shifts - so 27.3% of this frame is invented canvas. The original sits at pixels 0,192 to 512,704. Compare it INSET by 40 pixels: over that core it differs from the original by only 5.301/255, the VAE round trip, but over the whole rectangle by 6.294/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of a printed invoice on a wooden desk beside a calculator and a pen, flat lay, soft daylight - the frame an outpainting tool is given, cropped from the published plate nss-d-invoice_00001_.png. Its companion expands it to 512x704 with 192 pixels added below alone, the one case where the origin does NOT move, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of a printed invoice on a wooden desk beside a calculator and a pen, flat lay, soft daylight, expanded to 512x704 - 192 pixels added below alone, the one case where the origin does NOT move - so 27.3% of this frame is invented canvas. The original sits at pixels 0,0 to 512,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 4.626/255, the VAE round trip, but over the whole rectangle by 5.717/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of a restaurant menu card standing on a laid table, cutlery and glassware, warm evening light - the frame an outpainting tool is given, cropped from the published plate nss-d-menu_00001_.png. Its companion expands it to 704x704 with 96 pixels added on all four sides, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of a restaurant menu card standing on a laid table, cutlery and glassware, warm evening light, expanded to 704x704 - 96 pixels added on all four sides - so 47.1% of this frame is invented canvas. The original sits at pixels 96,96 to 608,608. Compare it INSET by 40 pixels: over that core it differs from the original by only 7.121/255, the VAE round trip, but over the whole rectangle by 12.793/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of a scatter of crumpled paper receipts on a kitchen table, overhead view, harsh afternoon sun and hard shadows - the frame an outpainting tool is given, cropped from the published plate nss-d-receipts_00001_.png. Its companion expands it to 768x640 with a different amount on every side, which is the case a tool assuming symmetry fails, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of a scatter of crumpled paper receipts on a kitchen table, overhead view, harsh afternoon sun and hard shadows, expanded to 768x640 - a different amount on every side, which is the case a tool assuming symmetry fails - so 46.7% of this frame is invented canvas. The original sits at pixels 64,32 to 576,544. Compare it INSET by 40 pixels: over that core it differs from the original by only 8.885/255, the VAE round trip, but over the whole rectangle by 15.578/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of rustic sourdough loaves on a wooden board, flour dust, warm bakery light - the frame an outpainting tool is given, cropped from the published plate nss-f-bread-rustic_00001_.png. Its companion expands it to 768x512 with 128 pixels added to the left and right, the symmetric case, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of rustic sourdough loaves on a wooden board, flour dust, warm bakery light, expanded to 768x512 - 128 pixels added to the left and right, the symmetric case - so 33.3% of this frame is invented canvas. The original sits at pixels 128,0 to 640,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 10.419/255, the VAE round trip, but over the whole rectangle by 12.644/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of close up of coffee being poured into a white cup, steam, cafe counter, shallow depth - the frame an outpainting tool is given, cropped from the published plate nss-f-coffee-pour_00001_.png. Its companion expands it to 512x768 with 128 pixels added above and below, the symmetric case turned on its side, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of close up of coffee being poured into a white cup, steam, cafe counter, shallow depth, expanded to 512x768 - 128 pixels added above and below, the symmetric case turned on its side - so 33.3% of this frame is invented canvas. The original sits at pixels 0,128 to 512,640. Compare it INSET by 40 pixels: over that core it differs from the original by only 4.288/255, the VAE round trip, but over the whole rectangle by 5.818/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of overhead plate of fresh pasta with herbs on a linen cloth, natural window light - the frame an outpainting tool is given, cropped from the published plate nss-f-plate-pasta_00001_.png. Its companion expands it to 704x512 with 192 pixels added to the right alone, so the original is no longer centred, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of overhead plate of fresh pasta with herbs on a linen cloth, natural window light, expanded to 704x512 - 192 pixels added to the right alone, so the original is no longer centred - so 27.3% of this frame is invented canvas. The original sits at pixels 0,0 to 512,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 8.016/255, the VAE round trip, but over the whole rectangle by 10.024/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of overhead colourful garden salad in a ceramic bowl, marble surface, soft daylight - the frame an outpainting tool is given, cropped from the published plate nss-f-plate-salad_00001_.png. Its companion expands it to 704x512 with 192 pixels added to the left alone, so the original's origin moves, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of overhead colourful garden salad in a ceramic bowl, marble surface, soft daylight, expanded to 704x512 - 192 pixels added to the left alone, so the original's origin moves - so 27.3% of this frame is invented canvas. The original sits at pixels 192,0 to 704,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 8.419/255, the VAE round trip, but over the whole rectangle by 11.066/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of stylised fantasy mage character render, full body, neutral grey backdrop - the frame an outpainting tool is given, cropped from the published plate nss-g-char-mage_00001_.png. Its companion expands it to 512x704 with 192 pixels added above alone, so every row index shifts, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of stylised fantasy mage character render, full body, neutral grey backdrop, expanded to 512x704 - 192 pixels added above alone, so every row index shifts - so 27.3% of this frame is invented canvas. The original sits at pixels 0,192 to 512,704. Compare it INSET by 40 pixels: over that core it differs from the original by only 5.477/255, the VAE round trip, but over the whole rectangle by 6.674/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of fantasy video game environment, stone keep on a cliff at sunset - the frame an outpainting tool is given, cropped from the published plate nss-g-fantasy-keep_00001_.png. Its companion expands it to 704x704 with 96 pixels added on all four sides, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of fantasy video game environment, stone keep on a cliff at sunset, expanded to 704x704 - 96 pixels added on all four sides - so 47.1% of this frame is invented canvas. The original sits at pixels 96,96 to 608,608. Compare it INSET by 40 pixels: over that core it differs from the original by only 7.895/255, the VAE round trip, but over the whole rectangle by 10.907/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of isometric stylised village diorama, soft shading, mobile style - the frame an outpainting tool is given, cropped from the published plate nss-g-isometric_00001_.png. Its companion expands it to 768x640 with a different amount on every side, which is the case a tool assuming symmetry fails, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of isometric stylised village diorama, soft shading, mobile style, expanded to 768x640 - a different amount on every side, which is the case a tool assuming symmetry fails - so 46.7% of this frame is invented canvas. The original sits at pixels 64,32 to 576,544. Compare it INSET by 40 pixels: over that core it differs from the original by only 8.719/255, the VAE round trip, but over the whole rectangle by 12.374/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of modern minimalist living room interior, large windows, natural daylight - the frame an outpainting tool is given, cropped from the published plate nss-l-interior-mod_00001_.png. Its companion expands it to 768x512 with 128 pixels added to the left and right, the symmetric case, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of modern minimalist living room interior, large windows, natural daylight, expanded to 768x512 - 128 pixels added to the left and right, the symmetric case - so 33.3% of this frame is invented canvas. The original sits at pixels 128,0 to 640,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 9.931/255, the VAE round trip, but over the whole rectangle by 13.41/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of an open paper accounts ledger with ruled columns on a desk, fountain pen resting on the page, window light - the frame an outpainting tool is given, cropped from the published plate nss-ledger-book_00001_.png. Its companion expands it to 512x768 with 128 pixels added above and below, the symmetric case turned on its side, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of an open paper accounts ledger with ruled columns on a desk, fountain pen resting on the page, window light, expanded to 512x768 - 128 pixels added above and below, the symmetric case turned on its side - so 33.3% of this frame is invented canvas. The original sits at pixels 0,128 to 512,640. Compare it INSET by 40 pixels: over that core it differs from the original by only 5.179/255, the VAE round trip, but over the whole rectangle by 6.829/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of tabby cat sitting on a windowsill, natural backlight, indoor - the frame an outpainting tool is given, cropped from the published plate nss-n-cat-window_00001_.png. Its companion expands it to 704x512 with 192 pixels added to the right alone, so the original is no longer centred, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of tabby cat sitting on a windowsill, natural backlight, indoor, expanded to 704x512 - 192 pixels added to the right alone, so the original is no longer centred - so 27.3% of this frame is invented canvas. The original sits at pixels 0,0 to 512,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 4.614/255, the VAE round trip, but over the whole rectangle by 5.838/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of middle aged indian man accountant at a desk with paper ledgers and a calculator, desk lamp - the frame an outpainting tool is given, cropped from the published plate nss-p-accountant_00001_.png. Its companion expands it to 704x512 with 192 pixels added to the left alone, so the original's origin moves, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of middle aged indian man accountant at a desk with paper ledgers and a calculator, desk lamp, expanded to 704x512 - 192 pixels added to the left alone, so the original's origin moves - so 27.3% of this frame is invented canvas. The original sits at pixels 192,0 to 704,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 3.731/255, the VAE round trip, but over the whole rectangle by 4.487/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of wheelchair user chef at a lowered prep counter slicing vegetables, professional kitchen - the frame an outpainting tool is given, cropped from the published plate nss-p-chef-wheel_00001_.png. Its companion expands it to 512x704 with 192 pixels added above alone, so every row index shifts, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of wheelchair user chef at a lowered prep counter slicing vegetables, professional kitchen, expanded to 512x704 - 192 pixels added above alone, so every row index shifts - so 27.3% of this frame is invented canvas. The original sits at pixels 0,192 to 512,704. Compare it INSET by 40 pixels: over that core it differs from the original by only 5.544/255, the VAE round trip, but over the whole rectangle by 6.677/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of older white man farmer crouching in a vegetable field holding produce, overcast daylight - the frame an outpainting tool is given, cropped from the published plate nss-p-farmer-field_00001_.png. Its companion expands it to 512x704 with 192 pixels added below alone, the one case where the origin does NOT move, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of older white man farmer crouching in a vegetable field holding produce, overcast daylight, expanded to 512x704 - 192 pixels added below alone, the one case where the origin does NOT move - so 27.3% of this frame is invented canvas. The original sits at pixels 0,0 to 512,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 4.112/255, the VAE round trip, but over the whole rectangle by 5.098/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of a South Asian woman electrician in work clothes and safety glasses checking a consumer unit, domestic hallway, tool bag open, natural light - the frame an outpainting tool is given, cropped from the published plate nss-p-trades-w_00001_.png. Its companion expands it to 704x704 with 96 pixels added on all four sides, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of a South Asian woman electrician in work clothes and safety glasses checking a consumer unit, domestic hallway, tool bag open, natural light, expanded to 704x704 - 96 pixels added on all four sides - so 47.1% of this frame is invented canvas. The original sits at pixels 96,96 to 608,608. Compare it INSET by 40 pixels: over that core it differs from the original by only 3.963/255, the VAE round trip, but over the whole rectangle by 10.812/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of glass water bottle on a plain white studio background, soft light - the frame an outpainting tool is given, cropped from the published plate nss-pr-bottle_00001_.png. Its companion expands it to 768x640 with a different amount on every side, which is the case a tool assuming symmetry fails, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of glass water bottle on a plain white studio background, soft light, expanded to 768x640 - a different amount on every side, which is the case a tool assuming symmetry fails - so 46.7% of this frame is invented canvas. The original sits at pixels 64,32 to 576,544. Compare it INSET by 40 pixels: over that core it differs from the original by only 6.819/255, the VAE round trip, but over the whole rectangle by 7.337/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of single running shoe on a plain white studio background, soft light - the frame an outpainting tool is given, cropped from the published plate nss-pr-shoe_00001_.png. Its companion expands it to 768x512 with 128 pixels added to the left and right, the symmetric case, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of single running shoe on a plain white studio background, soft light, expanded to 768x512 - 128 pixels added to the left and right, the symmetric case - so 33.3% of this frame is invented canvas. The original sits at pixels 128,0 to 640,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 9.491/255, the VAE round trip, but over the whole rectangle by 10.432/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of empty cafe interior before opening, chairs down, morning sun across tables - the frame an outpainting tool is given, cropped from the published plate nss-s-cafe-empty_00001_.png. Its companion expands it to 512x768 with 128 pixels added above and below, the symmetric case turned on its side, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of empty cafe interior before opening, chairs down, morning sun across tables, expanded to 512x768 - 128 pixels added above and below, the symmetric case turned on its side - so 33.3% of this frame is invented canvas. The original sits at pixels 0,128 to 512,640. Compare it INSET by 40 pixels: over that core it differs from the original by only 4.615/255, the VAE round trip, but over the whole rectangle by 8.693/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of a cluttered joinery workshop bench with hand tools and sawdust, window light from the left - the frame an outpainting tool is given, cropped from the published plate nss-s-workshop_00001_.png. Its companion expands it to 704x512 with 192 pixels added to the right alone, so the original is no longer centred, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of a cluttered joinery workshop bench with hand tools and sawdust, window light from the left, expanded to 704x512 - 192 pixels added to the right alone, so the original is no longer centred - so 27.3% of this frame is invented canvas. The original sits at pixels 0,0 to 512,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 6.862/255, the VAE round trip, but over the whole rectangle by 7.908/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of person lifting weights in a gym, indoor lighting - the frame an outpainting tool is given, cropped from the published plate nss-sp-gym_00001_.png. Its companion expands it to 704x512 with 192 pixels added to the left alone, so the original's origin moves, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of person lifting weights in a gym, indoor lighting, expanded to 704x512 - 192 pixels added to the left alone, so the original's origin moves - so 27.3% of this frame is invented canvas. The original sits at pixels 192,0 to 704,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 2.509/255, the VAE round trip, but over the whole rectangle by 4.2/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of abstract warm bokeh lights out of focus, dark background - the frame an outpainting tool is given, cropped from the published plate nss-t-bokeh-warm_00001_.png. Its companion expands it to 512x704 with 192 pixels added above alone, so every row index shifts, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of abstract warm bokeh lights out of focus, dark background, expanded to 512x704 - 192 pixels added above alone, so every row index shifts - so 27.3% of this frame is invented canvas. The original sits at pixels 0,192 to 512,704. Compare it INSET by 40 pixels: over that core it differs from the original by only 3.972/255, the VAE round trip, but over the whole rectangle by 5.543/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of flat red brick wall - the frame an outpainting tool is given, cropped from the published plate nss-t-tex-brick_00001_.png. Its companion expands it to 512x704 with 192 pixels added below alone, the one case where the origin does NOT move, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of flat red brick wall, expanded to 512x704 - 192 pixels added below alone, the one case where the origin does NOT move - so 27.3% of this frame is invented canvas. The original sits at pixels 0,0 to 512,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 5.876/255, the VAE round trip, but over the whole rectangle by 7.834/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of close up woven linen fabric texture, natural fibres - the frame an outpainting tool is given, cropped from the published plate nss-t-tex-linen_00001_.png. Its companion expands it to 704x704 with 96 pixels added on all four sides, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of close up woven linen fabric texture, natural fibres, expanded to 704x704 - 96 pixels added on all four sides - so 47.1% of this frame is invented canvas. The original sits at pixels 96,96 to 608,608. Compare it INSET by 40 pixels: over that core it differs from the original by only 14.966/255, the VAE round trip, but over the whole rectangle by 16.919/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of sports motorcycle parked at a kerb, sunny day - the frame an outpainting tool is given, cropped from the published plate nss-v-moto-sport_00001_.png. Its companion expands it to 768x512 with 128 pixels added to the left and right, the symmetric case, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of sports motorcycle parked at a kerb, sunny day, expanded to 768x512 - 128 pixels added to the left and right, the symmetric case - so 33.3% of this frame is invented canvas. The original sits at pixels 128,0 to 640,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 8.192/255, the VAE round trip, but over the whole rectangle by 10.954/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of green farm tractor in a field, blue sky - the frame an outpainting tool is given, cropped from the published plate nss-v-tractor_00001_.png. Its companion expands it to 512x768 with 128 pixels added above and below, the symmetric case turned on its side, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of green farm tractor in a field, blue sky, expanded to 512x768 - 128 pixels added above and below, the symmetric case turned on its side - so 33.3% of this frame is invented canvas. The original sits at pixels 0,128 to 512,640. Compare it INSET by 40 pixels: over that core it differs from the original by only 8.434/255, the VAE round trip, but over the whole rectangle by 14.105/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of a warehouse worker holding a clipboard beside tall racking, high bay lighting, boxes on pallets - the frame an outpainting tool is given, cropped from the published plate nss-warehouse-clip_00001_.png. Its companion expands it to 704x512 with 192 pixels added to the right alone, so the original is no longer centred, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of a warehouse worker holding a clipboard beside tall racking, high bay lighting, boxes on pallets, expanded to 704x512 - 192 pixels added to the right alone, so the original is no longer centred - so 27.3% of this frame is invented canvas. The original sits at pixels 0,0 to 512,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 6.222/255, the VAE round trip, but over the whole rectangle by 7.652/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of colourful stylised platformer level, floating islands, bright palette - the frame an outpainting tool is given, cropped from the published plate nss-g-platformer_00001_.png. Its companion expands it to 704x512 with 192 pixels added to the left alone, so the original's origin moves, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of colourful stylised platformer level, floating islands, bright palette, expanded to 704x512 - 192 pixels added to the left alone, so the original's origin moves - so 27.3% of this frame is invented canvas. The original sits at pixels 192,0 to 704,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 7.477/255, the VAE round trip, but over the whole rectangle by 8.353/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of a printed invoice on a wooden desk beside a calculator and a pen, flat lay, soft daylight - the frame an outpainting tool is given, cropped from the published plate nss-invoice-desk_00001_.png. Its companion expands it to 512x704 with 192 pixels added above alone, so every row index shifts, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of a printed invoice on a wooden desk beside a calculator and a pen, flat lay, soft daylight, expanded to 512x704 - 192 pixels added above alone, so every row index shifts - so 27.3% of this frame is invented canvas. The original sits at pixels 0,192 to 512,704. Compare it INSET by 40 pixels: over that core it differs from the original by only 6.206/255, the VAE round trip, but over the whole rectangle by 8.328/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of path through a deciduous forest, dappled sunlight, receding depth - the frame an outpainting tool is given, cropped from the published plate nss-l-forest-path_00001_.png. Its companion expands it to 512x704 with 192 pixels added below alone, the one case where the origin does NOT move, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of path through a deciduous forest, dappled sunlight, receding depth, expanded to 512x704 - 192 pixels added below alone, the one case where the origin does NOT move - so 27.3% of this frame is invented canvas. The original sits at pixels 0,0 to 512,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 10.599/255, the VAE round trip, but over the whole rectangle by 11.516/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of small garden birds on a wooden feeder, natural daylight - the frame an outpainting tool is given, cropped from the published plate nss-n-birds-feed_00001_.png. Its companion expands it to 704x704 with 96 pixels added on all four sides, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of small garden birds on a wooden feeder, natural daylight, expanded to 704x704 - 96 pixels added on all four sides - so 47.1% of this frame is invented canvas. The original sits at pixels 96,96 to 608,608. Compare it INSET by 40 pixels: over that core it differs from the original by only 5.083/255, the VAE round trip, but over the whole rectangle by 6.785/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of brown horse standing in a field, overcast daylight - the frame an outpainting tool is given, cropped from the published plate nss-n-horse-field_00001_.png. Its companion expands it to 768x640 with a different amount on every side, which is the case a tool assuming symmetry fails, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of brown horse standing in a field, overcast daylight, expanded to 768x640 - a different amount on every side, which is the case a tool assuming symmetry fails - so 46.7% of this frame is invented canvas. The original sits at pixels 64,32 to 576,544. Compare it INSET by 40 pixels: over that core it differs from the original by only 5.811/255, the VAE round trip, but over the whole rectangle by 9.977/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of a market stall vendor arranging fresh vegetables on his stall, outdoor daylight, crates and price cards, documentary - the frame an outpainting tool is given, cropped from the published plate nss-owner-market_00001_.png. Its companion expands it to 768x512 with 128 pixels added to the left and right, the symmetric case, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of a market stall vendor arranging fresh vegetables on his stall, outdoor daylight, crates and price cards, documentary, expanded to 768x512 - 128 pixels added to the left and right, the symmetric case - so 33.3% of this frame is invented canvas. The original sits at pixels 128,0 to 640,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 10.987/255, the VAE round trip, but over the whole rectangle by 12.497/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of a delivery driver with a parcel and a handheld scanner at a residential door, early evening light, van blurred behind - the frame an outpainting tool is given, cropped from the published plate nss-p-driver-eve_00001_.png. Its companion expands it to 512x768 with 128 pixels added above and below, the symmetric case turned on its side, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of a delivery driver with a parcel and a handheld scanner at a residential door, early evening light, van blurred behind, expanded to 512x768 - 128 pixels added above and below, the symmetric case turned on its side - so 33.3% of this frame is invented canvas. The original sits at pixels 0,128 to 512,640. Compare it INSET by 40 pixels: over that core it differs from the original by only 5.174/255, the VAE round trip, but over the whole rectangle by 7.379/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of a Black man in his seventies with grey hair arranging vegetables at an outdoor market stall, overcast daylight, crates and handwritten price cards - the frame an outpainting tool is given, cropped from the published plate nss-p-market-sen_00001_.png. Its companion expands it to 704x512 with 192 pixels added to the right alone, so the original is no longer centred, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of a Black man in his seventies with grey hair arranging vegetables at an outdoor market stall, overcast daylight, crates and handwritten price cards, expanded to 704x512 - 192 pixels added to the right alone, so the original is no longer centred - so 27.3% of this frame is invented canvas. The original sits at pixels 0,0 to 512,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 4.126/255, the VAE round trip, but over the whole rectangle by 6.273/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of a Black woman hairdresser in her forties in a bright salon, mirrors and styling chairs behind, professional, looking at camera - the frame an outpainting tool is given, cropped from the published plate nss-p-salon-mid_00001_.png. Its companion expands it to 704x512 with 192 pixels added to the left alone, so the original's origin moves, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of a Black woman hairdresser in her forties in a bright salon, mirrors and styling chairs behind, professional, looking at camera, expanded to 704x512 - 192 pixels added to the left alone, so the original's origin moves - so 27.3% of this frame is invented canvas. The original sits at pixels 192,0 to 704,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 4.042/255, the VAE round trip, but over the whole rectangle by 4.466/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of potted houseplant on a plain neutral background, soft daylight - the frame an outpainting tool is given, cropped from the published plate nss-pr-plant-pot_00001_.png. Its companion expands it to 512x704 with 192 pixels added above alone, so every row index shifts, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of potted houseplant on a plain neutral background, soft daylight, expanded to 512x704 - 192 pixels added above alone, so every row index shifts - so 27.3% of this frame is invented canvas. The original sits at pixels 0,192 to 512,704. Compare it INSET by 40 pixels: over that core it differs from the original by only 3.898/255, the VAE round trip, but over the whole rectangle by 5.136/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
A 512x512 view of the exterior of a small independent bakery on a quiet street, awning and window display, morning light - the frame an outpainting tool is given, cropped from the published plate nss-s-bakery_00001_.png. Its companion expands it to 512x704 with 192 pixels added below alone, the one case where the origin does NOT move, and the boundary record in this group states the exact rectangle this image occupies inside that frame - so where it ended up can be checked rather than eyeballed.
The same view of the exterior of a small independent bakery on a quiet street, awning and window display, morning light, expanded to 512x704 - 192 pixels added below alone, the one case where the origin does NOT move - so 27.3% of this frame is invented canvas. The original sits at pixels 0,0 to 512,512. Compare it INSET by 40 pixels: over that core it differs from the original by only 6.452/255, the VAE round trip, but over the whole rectangle by 7.26/255, because the pad deliberately feathers the original's outer edge into the new area. A test that expects the whole rectangle untouched fails on a correct tool.
Where the original view of head and shoulders portrait of an older white woman with short grey hair, plain cream background landed inside the expanded frame, as JSON: the rectangle 128,0 to 640,512 in a 768x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of head and shoulders portrait of an east asian woman with glasses, plain warm grey background landed inside the expanded frame, as JSON: the rectangle 0,128 to 512,640 in a 512x768 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of head and shoulders portrait of a bearded white man in his thirties, plain olive studio background landed inside the expanded frame, as JSON: the rectangle 0,0 to 512,512 in a 704x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of head and shoulders portrait of a south asian man in his fifties, plain slate studio background landed inside the expanded frame, as JSON: the rectangle 192,0 to 704,512 in a 704x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of head and shoulders portrait of a smiling white woman against a plain blue studio background landed inside the expanded frame, as JSON: the rectangle 0,0 to 512,512 in a 512x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of head and shoulders portrait of a young hijabi woman against a plain neutral studio background landed inside the expanded frame, as JSON: the rectangle 96,96 to 608,608 in a 704x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of head and shoulders portrait of an older east asian man against a plain warm studio background landed inside the expanded frame, as JSON: the rectangle 64,32 to 576,544 in a 768x640 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of abstract album cover artwork, bold overlapping colour shapes, grainy print texture landed inside the expanded frame, as JSON: the rectangle 128,0 to 640,512 in a 768x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of retro album cover artwork, sunset gradient and geometric lines, 1980s print feel landed inside the expanded frame, as JSON: the rectangle 0,128 to 512,640 in a 512x768 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of book cover artwork, weathered antique map and brass instruments on wood landed inside the expanded frame, as JSON: the rectangle 0,0 to 512,512 in a 704x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of book cover artwork, misty mountain ridge at dawn, layered depth, serene landed inside the expanded frame, as JSON: the rectangle 192,0 to 704,512 in a 704x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of a delivery note on a clipboard held against a van door, outdoor overcast light, slight angle landed inside the expanded frame, as JSON: the rectangle 0,192 to 512,704 in a 512x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of a printed invoice on a wooden desk beside a calculator and a pen, flat lay, soft daylight landed inside the expanded frame, as JSON: the rectangle 0,0 to 512,512 in a 512x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of a restaurant menu card standing on a laid table, cutlery and glassware, warm evening light landed inside the expanded frame, as JSON: the rectangle 96,96 to 608,608 in a 704x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of a scatter of crumpled paper receipts on a kitchen table, overhead view, harsh afternoon sun and hard shadows landed inside the expanded frame, as JSON: the rectangle 64,32 to 576,544 in a 768x640 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of rustic sourdough loaves on a wooden board, flour dust, warm bakery light landed inside the expanded frame, as JSON: the rectangle 128,0 to 640,512 in a 768x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of close up of coffee being poured into a white cup, steam, cafe counter, shallow depth landed inside the expanded frame, as JSON: the rectangle 0,128 to 512,640 in a 512x768 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of overhead plate of fresh pasta with herbs on a linen cloth, natural window light landed inside the expanded frame, as JSON: the rectangle 0,0 to 512,512 in a 704x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of overhead colourful garden salad in a ceramic bowl, marble surface, soft daylight landed inside the expanded frame, as JSON: the rectangle 192,0 to 704,512 in a 704x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of stylised fantasy mage character render, full body, neutral grey backdrop landed inside the expanded frame, as JSON: the rectangle 0,192 to 512,704 in a 512x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of fantasy video game environment, stone keep on a cliff at sunset landed inside the expanded frame, as JSON: the rectangle 96,96 to 608,608 in a 704x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of isometric stylised village diorama, soft shading, mobile style landed inside the expanded frame, as JSON: the rectangle 64,32 to 576,544 in a 768x640 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of modern minimalist living room interior, large windows, natural daylight landed inside the expanded frame, as JSON: the rectangle 128,0 to 640,512 in a 768x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of an open paper accounts ledger with ruled columns on a desk, fountain pen resting on the page, window light landed inside the expanded frame, as JSON: the rectangle 0,128 to 512,640 in a 512x768 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of tabby cat sitting on a windowsill, natural backlight, indoor landed inside the expanded frame, as JSON: the rectangle 0,0 to 512,512 in a 704x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of middle aged indian man accountant at a desk with paper ledgers and a calculator, desk lamp landed inside the expanded frame, as JSON: the rectangle 192,0 to 704,512 in a 704x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of wheelchair user chef at a lowered prep counter slicing vegetables, professional kitchen landed inside the expanded frame, as JSON: the rectangle 0,192 to 512,704 in a 512x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of older white man farmer crouching in a vegetable field holding produce, overcast daylight landed inside the expanded frame, as JSON: the rectangle 0,0 to 512,512 in a 512x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of a South Asian woman electrician in work clothes and safety glasses checking a consumer unit, domestic hallway, tool bag open, natural light landed inside the expanded frame, as JSON: the rectangle 96,96 to 608,608 in a 704x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of glass water bottle on a plain white studio background, soft light landed inside the expanded frame, as JSON: the rectangle 64,32 to 576,544 in a 768x640 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of single running shoe on a plain white studio background, soft light landed inside the expanded frame, as JSON: the rectangle 128,0 to 640,512 in a 768x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of empty cafe interior before opening, chairs down, morning sun across tables landed inside the expanded frame, as JSON: the rectangle 0,128 to 512,640 in a 512x768 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of a cluttered joinery workshop bench with hand tools and sawdust, window light from the left landed inside the expanded frame, as JSON: the rectangle 0,0 to 512,512 in a 704x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of person lifting weights in a gym, indoor lighting landed inside the expanded frame, as JSON: the rectangle 192,0 to 704,512 in a 704x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of abstract warm bokeh lights out of focus, dark background landed inside the expanded frame, as JSON: the rectangle 0,192 to 512,704 in a 512x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of flat red brick wall landed inside the expanded frame, as JSON: the rectangle 0,0 to 512,512 in a 512x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of close up woven linen fabric texture, natural fibres landed inside the expanded frame, as JSON: the rectangle 96,96 to 608,608 in a 704x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of sports motorcycle parked at a kerb, sunny day landed inside the expanded frame, as JSON: the rectangle 128,0 to 640,512 in a 768x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of green farm tractor in a field, blue sky landed inside the expanded frame, as JSON: the rectangle 0,128 to 512,640 in a 512x768 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of a warehouse worker holding a clipboard beside tall racking, high bay lighting, boxes on pallets landed inside the expanded frame, as JSON: the rectangle 0,0 to 512,512 in a 704x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of colourful stylised platformer level, floating islands, bright palette landed inside the expanded frame, as JSON: the rectangle 192,0 to 704,512 in a 704x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of a printed invoice on a wooden desk beside a calculator and a pen, flat lay, soft daylight landed inside the expanded frame, as JSON: the rectangle 0,192 to 512,704 in a 512x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of path through a deciduous forest, dappled sunlight, receding depth landed inside the expanded frame, as JSON: the rectangle 0,0 to 512,512 in a 512x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of small garden birds on a wooden feeder, natural daylight landed inside the expanded frame, as JSON: the rectangle 96,96 to 608,608 in a 704x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of brown horse standing in a field, overcast daylight landed inside the expanded frame, as JSON: the rectangle 64,32 to 576,544 in a 768x640 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of a market stall vendor arranging fresh vegetables on his stall, outdoor daylight, crates and price cards, documentary landed inside the expanded frame, as JSON: the rectangle 128,0 to 640,512 in a 768x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of a delivery driver with a parcel and a handheld scanner at a residential door, early evening light, van blurred behind landed inside the expanded frame, as JSON: the rectangle 0,128 to 512,640 in a 512x768 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of a Black man in his seventies with grey hair arranging vegetables at an outdoor market stall, overcast daylight, crates and handwritten price cards landed inside the expanded frame, as JSON: the rectangle 0,0 to 512,512 in a 704x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of a Black woman hairdresser in her forties in a bright salon, mirrors and styling chairs behind, professional, looking at camera landed inside the expanded frame, as JSON: the rectangle 192,0 to 704,512 in a 704x512 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of potted houseplant on a plain neutral background, soft daylight landed inside the expanded frame, as JSON: the rectangle 0,192 to 512,704 in a 512x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
Where the original view of the exterior of a small independent bakery on a quiet street, awning and window display, morning light landed inside the expanded frame, as JSON: the rectangle 0,0 to 512,512 in a 512x704 canvas, the padding that produced it, and the 40-pixel feather that means the comparison must be inset. Authored BEFORE the expansion ran, so it is ground truth rather than a rectangle recovered afterwards by matching the original against the result - which is what a consumer would otherwise have to do, and which cannot distinguish a correct offset from a plausible one.
